Robotic Fabrication of Curved Folds


Between 2011 and 2014, I collaborated with RoboFold, a robotic fabrication company based in London, to develop processes to fold sheet metal with 6-axis robots. These techniques have applications in furniture & product design, sculpture and architecture. As a result we collaborated on curved folding projects with architects, engineers, facade designers and artists.

We periodically ran workshops with students, such as Shape to Fabrication 2011, which enabled us to share the processes at the same time as refining them.


The design process is a very intuitive one, combining both hands-on experimentation and digital simulation & fabrication.

We began by building lots of rough card models by hand, just to get a feel for how the card behaves physically and in order to generate lots of varied possible forms. These can be made by sketching on card sheets and folding along the ridge lines created, or just by running a blunt scalpel ‘freehand’ across the card, and seeing how the resulting form folds (or doesn’t!). These experiments quickly reveal the few underlying ‘rules’ to curved folds in sheets material – such as if you have two concave folds next to one another, one must be a ‘mountain’ fold and one must be a ‘valley’ fold, and so on. All of these rules are in reality just a result of the material remaining developable throughout the fold process – so while the entire sheet might appear to have a very complex form, all the individual sections that comprise it are either cylindrical, conic, tangent developable or (in some unusual cases) planar.

We then isolated a few of the most promising panel designs based on criteria such as the desired aesthetic, ease of folding, numbers of contact points required (never more than the number of robots we own!), panel size and nesting during transportation, desired overlap (if it is external cladding), and so on. These selected designs are then flattened and scanned into a computer, and redrawn as a vector-based template. At this point we have to make some educated guesses about where the rulings sit on each piece of the surface, and draw these on the surface. Finding the rulings is straightforward for some folds, but in others the surfaces might blend from one developable type to another (e.g. the rulings in one section might run to a point, suggesting that the surface is a conic, before becoming parallel as the surface becomes cylindrical). At the moment RoboFold are working on several processes to make this analysis easier.

Once the final flat pattern is refined, we send it to a lasercutter or cnc router and produce the template for the object in sheet metal. Each of the mountain and valley fold lines is either subtly scored along, or we use a dashed ‘perforated’ cut along the fold lines. The depth of the cut, or the use of perforations, depends on the particular fold we want to achieve.

At the same time we import the pattern file into Rhino and GrassHopper, and use the Kangaroo Plugin to digitally simulate the material characteristics of the sheet metal we will use for the final object. We treat each edge and ruling in the pattern as a rigid spring and give them an identical start and rest length, meaning during the Kangaroo physics simulation they will act like a network of rigid wires. In this way the simulation can approximate a non-stretchy material (i.e. like sheet metal or paper, rather than say rubber). However each face must also stay planar – this is not a problem with the triangular faces in the pattern, but for the quad faces we must also extract diagonals and add these to the network of springs. After this we select certain ‘hinge’ points in the pattern, and define what angle we want them to fold to. Because the patterns have only a single degree of freedom, as one crease is folded all others should fold simultaneously.

Finally, we run the fold simulation and record the positional (X,Y,Z) and normal data for certain points on the surface that we want the robot arms to ‘grasp’ during the fold. Through an inverse kinematics process this positional information is then used to locate the robotic manufacturing arms during production. Other details, such as when to turn on and off the vacuum effectors at the end of each arm, are added after the simulation has been run and ‘recorded’.

The entire process is quite an iterative one, so that sometimes a pattern is simulated in Kangaroo and as a result we go back to the original pattern and adjust it. Similarly we often fold one of the full-scale metal templates by hand before sending it to the robots. This gives the designer a feel for how the material is behaving during the process, and it might tell us that the robot arms need to be repositioned on the surface, or that the score lines need to be deeper.

Personally, I really value the process for being fundamentally driven by the material constraints of the material – the metals – used. This seems to be something rather rare in design at the moment, when everyone is producing textureless 3D printed models.


Process Videos


RoboFold Facade Fabrication Workshop September 2011

Simulation of curved folding – in sheet material, using Kangaroo and Lobster IK. These processes are applied to a table prototype developed by Gregory Epps and Robofold. The finished 1:2 scale card mock-up of the table was exhibited at the DesignLondon talks at the RCA, May 2011.


Various Images of the Process


Parametric effector model

Facade panel test pieces

‘Generative Design for Robotic Manufacture’ Workshop 2011

1:2 scale white card furniture mock-up

Robofold and furniture prototypes featured in MARK magazine Oct 2011